Self-Driving Car Engineer Nanodegree¶
Deep Learning¶
Project: Build a Traffic Sign Recognition Classifier¶
In this notebook, a template is provided for you to implement your functionality in stages which is required to successfully complete this project. If additional code is required that cannot be included in the notebook, be sure that the Python code is successfully imported and included in your submission, if necessary. Sections that begin with 'Implementation' in the header indicate where you should begin your implementation for your project. Note that some sections of implementation are optional, and will be marked with 'Optional' in the header.
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. In addition, Markdown cells can be edited by typically double-clicking the cell to enter edit mode.
# common import for all cells
SEED=202
DATA_DIR=r'/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/submission(notebook+html)/002/data'
OUT_DIR =r'/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/submission(notebook+html)/002/out'
# standard libs
import pickle
import csv
from timeit import default_timer as timer
import os
import sys
#visualisation
%matplotlib inline
import matplotlib.pyplot as plt
from IPython.display import Image
from IPython.display import display
# numerical libs
import cv2
import math
import random
import numpy as np
random.seed(SEED)
np.random.seed(SEED)
import tensorflow as tf
tf.set_random_seed(SEED)
from tensorflow.python.training import moving_averages
from tensorflow.contrib.framework import add_model_variable
sess = tf.InteractiveSession()
Step 0: Load The Data¶
# Load pickled data
# TODO: Fill this in based on where you saved the training and testing data
def load_data():
training_file = DATA_DIR + '/train.p'
testing_file = DATA_DIR + '/test.p'
classname_file = DATA_DIR + '/signnames.csv'
classnames = []
with open(classname_file) as _f:
rows = csv.reader(_f, delimiter=',')
next(rows, None) # skip the headers
for i, row in enumerate(rows):
assert(i==int(row[0]))
classnames.append(row[1])
with open(training_file, mode='rb') as f:
train = pickle.load(f)
with open(testing_file, mode='rb') as f:
test = pickle.load(f)
X_train, y_train = train['features'], train['labels']
X_test, y_test = test['features'], test['labels']
X_train = X_train.astype(np.float32)
y_train = y_train.astype(np.int32)
X_test = X_test.astype(np.float32)
y_test = y_test.astype(np.int32)
return classnames, X_train, y_train, X_test, y_test
Step 1: Dataset Summary & Exploration¶
The pickled data is a dictionary with 4 key/value pairs:
'features'is a 4D array containing raw pixel data of the traffic sign images, (num examples, width, height, channels).'labels'is a 1D array containing the label/class id of the traffic sign. The filesignnames.csvcontains id -> name mappings for each id.'sizes'is a list containing tuples, (width, height) representing the the original width and height the image.'coords'is a list containing tuples, (x1, y1, x2, y2) representing coordinates of a bounding box around the sign in the image. THESE COORDINATES ASSUME THE ORIGINAL IMAGE. THE PICKLED DATA CONTAINS RESIZED VERSIONS (32 by 32) OF THESE IMAGES
Complete the basic data summary below.
### Replace each question mark with the appropriate value.
classnames, X_train, y_train, X_test, y_test = load_data()
# TODO: Number of training examples
num_train = len(X_train)
# TODO: Number of testing examples.
num_test = len(X_test)
# TODO: What's the shape of an traffic sign image?
_, height, width, channel = X_train.shape
image_shape = (height, width, channel)
# TODO: How many unique classes/labels there are in the dataset.
num_class = len(np.unique(y_train))
print("Number of training examples =", num_train )
print("Number of testing examples =", num_test )
print("Image data shape =", image_shape)
print("Number of classes =", num_class)
Visualize the German Traffic Signs Dataset using the pickled file(s). This is open ended, suggestions include: plotting traffic sign images, plotting the count of each sign, etc.
The Matplotlib examples and gallery pages are a great resource for doing visualizations in Python.
NOTE: It's recommended you start with something simple first. If you wish to do more, come back to it after you've completed the rest of the sections.
### Data exploration visualization goes here.
### Feel free to use as many code cells as needed.
#help function to draw graph,etc
def get_label_image(c):
img=cv2.imread(DATA_DIR+'/signnames_all.jpg',1)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
H, W, _ = img.shape
dH = H/7.
dW = W/7.105
y = c//7
x = c%7
label_image = img[round(y*dH):round(y*dH+dH), round(x*dW):round(x*dW+dW),:]
label_image = cv2.resize(label_image, (0,0), fx=32./dW, fy=32./dH,)
return label_image
def insert_subimage(image, sub_image, y, x):
h, w, c = sub_image.shape
image[y:y+h, x:x+w, :]=sub_image
return image
images, labels = X_train, y_train
#images, labels = X_test, y_test
#count
#h = np.histogram(train_labels, bins=np.arange(num_class))
#results image
num_sample=10
results_image = 255.*np.ones(shape=(num_class*height,(num_sample+2+22)*width, channel),dtype=np.float32)
for c in range(num_class):
label_image = get_label_image(c)
insert_subimage(results_image, label_image, c*height, 0)
#make mean
idx = list(np.where(labels== c)[0])
mean_image = np.average(images[idx], axis=0)
insert_subimage(results_image, mean_image, c*height, width)
# imshow('mean_image',mean_image)
# imshow('label_image',label_image)
# cv2.waitKey(0)
#make random sample
for n in range(num_sample):
sample_image = images[np.random.choice(idx)]
insert_subimage(results_image, sample_image, c*height, (2+n)*width)
#print summary
count=len(idx)
percentage = float(count)/float(len(images))
cv2.putText(results_image, '%02d:%-6s'%(c, classnames[c]), ((2+num_sample)*width, int((c+0.7)*height)),cv2.FONT_HERSHEY_SIMPLEX,0.5,(0,0,0),1)
cv2.putText(results_image, '[%4d]'%(count), ((2+num_sample+14)*width, int((c+0.7)*height)),cv2.FONT_HERSHEY_SIMPLEX,0.5,(0,0,255),1)
cv2.rectangle(results_image,((2+num_sample+16)*width, c*height),((2+num_sample+16)*width + round(percentage*3000), (c+1)*height),(0,0,255),-1)
cv2.imwrite(OUT_DIR+'/train_data_summary.jpg',cv2.cvtColor(results_image, cv2.COLOR_BGR2RGB))
#cv2.imwrite(OUT_DIR+'/test_data_summary.jpg',cv2.cvtColor(results_image, cv2.COLOR_BGR2RGB))
#imshow('results_image',results_image)
#cv2.waitKey(0)
print('** training data summary **')
print('\t1st column: label(image)')
print('\t2nd column: mean image')
print('\tother column: example images')
print('\tblack text: label')
print('\tblue text: sanple count for each class and histogram plot')
plt.rcParams["figure.figsize"] = (25,25)
plt.imshow(results_image.astype(np.uint8))
plt.axis('off')
plt.show()
comments
It is noted that the mean image of each class is rather well define. Hence we do not expect the classification to be too difficult.The data varies greatly in illumintion. There is some affine transformation and blur. Lastly, the histogram of the samples shows that the class is not balanced.
Step 2: Design and Test a Model Architecture¶
Design and implement a deep learning model that learns to recognize traffic signs. Train and test your model on the German Traffic Sign Dataset.
There are various aspects to consider when thinking about this problem:
- Neural network architecture
- Play around preprocessing techniques (normalization, rgb to grayscale, etc)
- Number of examples per label (some have more than others).
- Generate fake data.
Here is an example of a published baseline model on this problem. It's not required to be familiar with the approach used in the paper but, it's good practice to try to read papers like these.
NOTE: The LeNet-5 implementation shown in the classroom at the end of the CNN lesson is a solid starting point. You'll have to change the number of classes and possibly the preprocessing, but aside from that it's plug and play!
Implementation¶
Use the code cell (or multiple code cells, if necessary) to implement the first step of your project. Once you have completed your implementation and are satisfied with the results, be sure to thoroughly answer the questions that follow.
### Preprocess the data here.
### Feel free to use as many code cells as needed.
# our preprocessing is built inside the network:
# with tf.variable_scope('preprocess') as scope:
# input = bn(input, name='b1')
# input = conv2d(input, num_kernels=8, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME', has_bias=True, name='c1')
# input = prelu(input, name='r1')
# input = conv2d(input, num_kernels=8, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', has_bias=True, name='c2')
# input = prelu(input, name='r2')
Question 1¶
Describe how you preprocessed the data. Why did you choose that technique?
Answer:
Note: you can refer to my github or blog for more details:)
- https://github.com/hengck23-udacity/udacity-driverless-car-nd-p2
- https://medium.com/@hengcherkeng/updated-my-99-40-solution-to-udacity-nanodegree-project-p2-traffic-sign-classification-5580ae5bd51f#.28l99uq77
We did not use any data preprocessing methods (e.g. color conversion, whitening, histograms). Instead, we use the first few layer layers of our network for preprocessing. This is inspired by [1]. Like the "spatial transformer", we want to design a "illumination transformer" to auto correct the image color, contrast, etc. We use small convolution kernels of 3x3 and 1x1 and learnable parametric ReLU.
[1]"Systematic evaluation of CNN advances on the ImageNet"-Dmytro Mishkin, Nikolay Sergievskiy, Jiri Matas, ARXIV 2016 (https://arxiv.org/abs/1606.02228)
### Generate additional data (OPTIONAL!)
### and split the data into training/validation/testing sets here.
### Feel free to use as many code cells as needed.
# split into train and validation.
def split_data(X_train, y_train, num_valid=3000):
# num_valid = num of validation samples
num = len(y_train) # 39209
index = list(range(num))
random.shuffle(index)
train_index=index[num_valid:]
valid_index=index[:num_valid]
train_images = X_train[train_index]
train_labels = y_train[train_index]
valid_images = X_train[valid_index]
valid_labels = y_train[valid_index]
return train_images, train_labels, valid_images, valid_labels
#this expand the train data by flipping.
# Note: this code is from : http://navoshta.com/traffic-signs-classification/ !!!!!!!!!!!!!!
# This is the only code *not* from me.
def extend_data_by_flipping(images, labels):
X=images
y=labels
# Classes of signs that, when flipped horizontally, should still be classified as the same class
self_flippable_horizontally = np.array([11, 12, 13, 15, 17, 18, 22, 26, 30, 35])
# Classes of signs that, when flipped vertically, should still be classified as the same class
self_flippable_vertically = np.array([1, 5, 12, 15, 17])
# Classes of signs that, when flipped horizontally and then vertically, should still be classified as the same class
self_flippable_both = np.array([32, 40])
# Classes of signs that, when flipped horizontally, would still be meaningful, but should be classified as some other class
cross_flippable = np.array([
[19, 20],
[33, 34],
[36, 37],
[38, 39],
[20, 19],
[34, 33],
[37, 36],
[39, 38],
])
num_classes = 43
X_extended = np.empty([0, X.shape[1], X.shape[2], X.shape[3]], dtype=np.float32)
y_extended = np.empty([0], dtype=np.int32)
for c in range(num_classes):
# First copy existing data for this class
X_extended = np.append(X_extended, X[y == c], axis=0)
# If we can flip images of this class horizontally and they would still belong to said class...
if c in self_flippable_horizontally:
# ...Copy their flipped versions into extended array.
X_extended = np.append(X_extended, X[y == c][:, :, ::-1, :], axis=0)
# If we can flip images of this class horizontally and they would belong to other class...
if c in cross_flippable[:, 0]:
# ...Copy flipped images of that other class to the extended array.
flip_class = cross_flippable[cross_flippable[:, 0] == c][0][1]
X_extended = np.append(X_extended, X[y == flip_class][:, :, ::-1, :], axis=0)
# Fill labels for added images set to current class.
y_extended = np.append(y_extended, np.full((X_extended.shape[0] - y_extended.shape[0]), c, dtype=np.int32))
# If we can flip images of this class vertically and they would still belong to said class...
if c in self_flippable_vertically:
# ...Copy their flipped versions into extended array.
X_extended = np.append(X_extended, X_extended[y_extended == c][:, ::-1, :, :], axis=0)
# Fill labels for added images set to current class.
y_extended = np.append(y_extended, np.full((X_extended.shape[0] - y_extended.shape[0]), c, dtype=np.int32))
# If we can flip images of this class horizontally AND vertically and they would still belong to said class...
if c in self_flippable_both:
# ...Copy their flipped versions into extended array.
X_extended = np.append(X_extended, X_extended[y_extended == c][:, ::-1, ::-1, :], axis=0)
# Fill labels for added images set to current class.
y_extended = np.append(y_extended, np.full((X_extended.shape[0] - y_extended.shape[0]), c, dtype=np.int32))
extend_datas = X_extended
extend_labels = y_extended
return (extend_datas, extend_labels)
# use opencv to do data agumentation
# see also: https://github.com/dmlc/mxnet/blob/master/python/mxnet/image.py
def perturb(image, keep, angle_limit=15, scale_limit=0.1, translate_limit=3, distort_limit=3, illumin_limit=0.7):
u=np.random.uniform()
if u>keep :
#geometric -------------
(W, H, C) = image.shape
center = np.array([W / 2., H / 2.])
da = np.random.uniform(low=-1, high=1) * angle_limit/180. * math.pi
scale = np.random.uniform(low=-1, high=1) * scale_limit + 1
cc = scale*math.cos(da)
ss = scale*math.sin(da)
rotation = np.array([[cc, ss],[-ss,cc]])
translation = np.random.uniform(low=-1, high=1, size=(1,2)) * translate_limit
distort = np.random.standard_normal(size=(4,2)) * distort_limit
pts1 = np.array([[0., 0.], [0., H], [W, H], [W, 0.]])
pts2 = np.matmul(pts1-center, rotation) + center + translation
#add perspective noise
pts2 = pts2 + distort
#http://milindapro.blogspot.jp/2015/05/opencv-filters-copymakeborder.html
matrix = cv2.getPerspectiveTransform(pts1.astype(np.float32), pts2.astype(np.float32))
perturb = cv2.warpPerspective(image, matrix, (W, H), flags=cv2.INTER_LINEAR,
borderMode=cv2.BORDER_REFLECT_101) # BORDER_WRAP #BORDER_REFLECT_101 #cv2.BORDER_CONSTANT BORDER_REPLICATE
#illumination -------------
#from mxnet code
#brightness
alpha = 1.0 + illumin_limit*random.uniform(-1, 1)
perturb *= alpha
perturb = np.clip(perturb,0.,255.)
pass
#contrast
coef = np.array([[[0.299, 0.587, 0.114]]]) #rgb to gray (YCbCr) : Y = 0.299R + 0.587G + 0.114B
alpha = 1.0 + illumin_limit*random.uniform(-1, 1)
gray = perturb * coef
gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
perturb *= alpha
perturb += gray
perturb = np.clip(perturb,0.,255.)
pass
#saturation
coef = np.array([[[0.299, 0.587, 0.114]]]) #rgb to gray (YCbCr) : Y = 0.299R + 0.587G + 0.114B
alpha = 1.0 + illumin_limit*random.uniform(-1, 1)
gray = perturb * coef
gray = np.sum(gray, axis=2, keepdims=True)
gray *= (1.0 - alpha)
perturb *= alpha
perturb += gray
perturb = np.clip(perturb,0.,255.)
pass
return perturb
else:
return image
def make_perturb_images(images, keep ):
arguments = np.zeros(images.shape)
for n in range(len(images)):
arguments[n] = perturb(images[n],keep = keep)
return arguments
# sample and shuffle the data such that each class has equal number of samples for training
def shuffle_data_uniform(datas, labels, num_class, num_per_class=None):
if num_per_class is None:
max_count = 0
for c in range(num_class):
idx = list(np.where(labels == c)[0])
count = len(idx)
max_count = max(count, max_count)
num_per_class = max_count
index = []
for c in range(num_class):
idx = list(np.where(labels == c)[0])
index = index + list(np.random.choice(idx, num_per_class))
random.shuffle(index)
shuffle_datas = datas[index]
shuffle_labels = labels[index]
return shuffle_datas, shuffle_labels
#generate next batch for sdg
def generate_train_batch_next(datas, labels, n, batch_size):
i = n*batch_size
batch_datas = datas [i:i+batch_size]
batch_labels = labels[i:i+batch_size]
return batch_datas, batch_labels
#prepare all data here
classnames, X_train, y_train, X_test, y_test = load_data()
train_images, train_labels, valid_images, valid_labels = split_data(X_train, y_train)
test_images, test_labels = X_test, y_test
num_train = len(train_images)
num_valid = len(valid_images)
num_test = len(test_images)
print('** some data setting **')
print('\theight, width, channel = %d, %d, %d'%(height, width, channel))
print('\tnum_test = %d'%num_test)
print('\tnum_valid = %d'%num_valid)
print('\tnum_train = %d'%num_train)
#train data filpping.
train_images, train_labels = extend_data_by_flipping(train_images, train_labels)
num_train_flip = len(train_images)
print('')
print('\tnum_train (after flip)= %d' % num_train_flip )
#train data augmentation
keep = 0.20 # 0.50 0.25 0.20 #0.15
num_per_class = 20000
num_argument = num_per_class*num_class
#argument_images, argument_labels = shuffle_data_uniform(train_images, train_labels, num_class, num_per_class=num_per_class)
#argument_images = make_perturb_images(argument_images, keep=keep)
#num_argument = len(argument_images)
print('\tnum_argument = %d' % num_argument)
#example of new generated train data --------------------------------------------
print('\n')
print('examples of argument images (first column is orginal)')
# results image
num_sample = 20
perturbance_per_sample = 20
results_image = 255. * np.ones(shape=(num_sample * height, (perturbance_per_sample+1)* width+2, channel),dtype=np.float32)
for j in range(num_sample):
i = random.randint(0, num_train_flip - 1)
image = train_images[i]
insert_subimage(results_image, image, j * height, 0)
for k in range(0, perturbance_per_sample):
perturb_image = perturb(image, keep=0)
insert_subimage(results_image, perturb_image, j*height, (k+1)*width+2)
cv2.imwrite(OUT_DIR+'/data_perturb.jpg',cv2.cvtColor(results_image, cv2.COLOR_BGR2RGB))
plt.rcParams["figure.figsize"] = (25,25)
plt.imshow(results_image.astype(np.uint8))
plt.axis('off')
plt.show()
Question 2¶
Describe how you set up the training, validation and testing data for your model. Optional: If you generated additional data, how did you generate the data? Why did you generate the data? What are the differences in the new dataset (with generated data) from the original dataset?
Answer: We set aside 3000 train samples for validation and the rest 36209 for training. For 12630 test data, we leave all of it untouched. In order to create more data, we use the methods
We first flip the train sample to extend the training set to 62259. This is inspired by the work of http://navoshta.com/traffic-signs-classification/
Next, to keep the class balance, we first re-sample the training set so that we have 20000 samples per class. This gives 43x2000=860000 train samples.
Finally, we randomly select 80% train samples to add random geometric transform and illumination transform. Geometric transform include rotate, scale, transalate and perspective distortion noise. Illumination transform includes brightness, contast and saturation. These transforms are selected to mimick the data variations we found in the training images, which we have shown above.
We note that from our experiments, the illumination transform gives good improvemet in the test accuracy. The new augmented data is more class balance and has more variations then the original data.
### Define your architecture here.
### Feel free to use as many code cells as needed.
# we define the basic ops first -----------------------------------------
## global varaiables ##
IS_TRAIN_PHASE = tf.placeholder(dtype=tf.bool, name='is_train_phase')
def conv2d(input, num_kernels=1, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', has_bias=True, name='conv'):
input_shape = input.get_shape().as_list()
assert len(input_shape)==4
C = input_shape[3]
H = kernel_size[0]
W = kernel_size[1]
K = num_kernels
##[filter_height, filter_width, in_channels, out_channels]
w = tf.get_variable(name=name+'_weight', shape=[H, W, C, K], initializer=tf.truncated_normal_initializer(stddev=0.1))
conv = tf.nn.conv2d(input, w, strides=stride, padding=padding, name=name)
if has_bias:
b = tf.get_variable(name=name + '_bias', shape=[K], initializer=tf.constant_initializer(0.0))
conv = conv+b
return conv
def relu(input, name='relu'):
act = tf.nn.relu(input, name=name)
return act
def prelu(input, name='prelu'):
alpha = tf.get_variable(name=name+'_alpha', shape=input.get_shape()[-1],
#initializer=tf.constant_initializer(0.25),
initializer=tf.random_uniform_initializer(minval=0.1, maxval=0.3),
dtype=tf.float32)
pos = tf.nn.relu(input)
neg = alpha * (input - abs(input)) * 0.5
return pos + neg
# very leaky relu
def vlrelu(input, alpha=0.25, name='vlrelu'): # alpha between 0.1 to 0.5
act =tf.maximum(alpha*input,input)
return act
def maxpool(input, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', has_bias=True, name='max' ):
H = kernel_size[0]
W = kernel_size[1]
pool = tf.nn.max_pool(input, ksize=[1, H, W, 1], strides=stride, padding=padding, name=name)
return pool
def avgpool(input, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', has_bias=True, is_global_pool=False, name='avg'):
if is_global_pool==True:
input_shape = input.get_shape().as_list()
assert len(input_shape) == 4
H = input_shape[1]
W = input_shape[2]
pool = tf.nn.avg_pool(input, ksize=[1, H, W, 1], strides=[1,H,W,1], padding='VALID', name=name)
pool = flatten(pool)
else:
H = kernel_size[0]
W = kernel_size[1]
pool = tf.nn.avg_pool(input, ksize=[1, H, W, 1], strides=stride, padding=padding, name=name)
return pool
def dropout(input, keep=1.0, name='drop'):
#drop = tf.cond(IS_TRAIN_PHASE, lambda: tf.nn.dropout(input, keep), lambda: input)
drop = tf.cond(IS_TRAIN_PHASE,
lambda: tf.nn.dropout(input, keep),
lambda: tf.nn.dropout(input, 1))
return drop
def flatten(input, name='flat'):
input_shape = input.get_shape().as_list() # list: [None, 9, 2]
dim = np.prod(input_shape[1:]) # dim = prod(9,2) = 18
flat = tf.reshape(input, [-1, dim], name=name) # -1 means "all"
return flat
def concat(input, name='cat'):
cat = tf.concat(concat_dim=3, values=input, name=name)
return cat
#https://github.com/tensorflow/tensorflow/blob/master/tensorflow/g3doc/api_docs/python/functions_and_classes/shard4/tf.contrib.layers.batch_norm.md
#http://www.bubufx.com/detail-1792794.html
def bn (input, decay=0.9, eps=1e-5, name='bn'):
with tf.variable_scope(name) as scope:
bn = tf.cond(IS_TRAIN_PHASE,
lambda: tf.contrib.layers.batch_norm(input, decay=decay, epsilon=eps, center=True, scale=True,
is_training=1,reuse=None,
updates_collections=None, scope=scope),
lambda: tf.contrib.layers.batch_norm(input, decay=decay, epsilon=eps, center=True, scale=True,
is_training=0, reuse=True,
updates_collections=None, scope=scope))
return bn
# basic building blocks
def conv2d_bn_relu(input, num_kernels=1, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='conv'):
with tf.variable_scope(name) as scope:
block = conv2d(input, num_kernels=num_kernels, kernel_size=kernel_size, stride=stride, padding=padding, has_bias=False)
block = bn(block)
block = relu(block)
return block
def bn_relu_conv2d (input, num_kernels=1, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='conv'):
with tf.variable_scope(name) as scope:
block = bn(input)
block = relu(block)
block = conv2d(block, num_kernels=num_kernels, kernel_size=kernel_size, stride=stride, padding=padding, has_bias=False)
return block
# modified dense block from the paper [1] "Densely Connected Convolutional Networks" - Gao Huang, Zhuang Liu, Kilian Q. Weinberger,
# Laurens van der Maaten, Arxiv 2016
# Modification:
# 1. the paper uses bn-relu-conv but we use conv-bn-relu
# 2. the paper uses dropout inside the block but we shift the dropout outside the block see network construction later
def dense_block_cbr (input, num=1, num_kernels=1, kernel_size=(1, 1), drop=None, name='DENSE'):
block = input
for n in range(num):
with tf.variable_scope(name+'_%d'%n) as scope:
conv = conv2d(block, num_kernels=num_kernels, kernel_size=kernel_size, stride=[1,1,1,1], padding='SAME', has_bias=False)
conv = bn(conv)
conv = relu(conv)
if drop is not None:
keep = (1 - drop) ** (1. / num)
conv = dropout(conv, keep=keep)
block = concat((block, conv))
return block
# the loss
def l2_regulariser(decay):
variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
for v in variables:
name = v.name
if 'weight' in name: #this is weight
l2 = decay * tf.nn.l2_loss(v)
tf.add_to_collection('losses', l2)
elif 'bias' in name: #this is bias
pass
elif 'beta' in name:
pass
elif 'gamma' in name:
pass
elif 'moving_mean' in name:
pass
elif 'moving_variance' in name:
pass
elif 'moments' in name:
pass
else:
#pass
#raise Exception('unknown variable type: %s ?'%name)
pass
l2_loss = tf.add_n(tf.get_collection('losses'))
return l2_loss
def cross_entropy(logit, label, name='cross_entropy'):
label = tf.cast(label, tf.int64)
cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logit, label), name=name)
return cross_entropy
def accuracy(prob, label, name='accuracy'):
correct_prediction = tf.equal(tf.argmax(prob, 1), tf.cast(label, tf.int64))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name=name)
return accuracy
# my densenet here!
#the inference part (without loss)
def DenseNet_3( input_shape=(1,1,1), output_shape = (1)):
H, W, C = input_shape
num_class = output_shape
input = tf.placeholder(shape=[None, H, W, C], dtype=tf.float32, name='input')
#color preprocessing using conv net:
#see "Systematic evaluation of CNN advances on the ImageNet"-Dmytro Mishkin, Nikolay Sergievskiy, Jiri Matas, ARXIV 2016
# https://arxiv.org/abs/1606.02228
# we use learnable prelu (different from paper) and 3x3 onv
with tf.variable_scope('preprocess') as scope:
input = bn(input, name='b1')
input = conv2d(input, num_kernels=8, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME', has_bias=True, name='c1')
input = prelu(input, name='r1')
input = conv2d(input, num_kernels=8, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', has_bias=True, name='c2')
input = prelu(input, name='r2')
with tf.variable_scope('block1') as scope:
block1 = conv2d_bn_relu(input, num_kernels=32, kernel_size=(5, 5), stride=[1, 1, 1, 1], padding='SAME')
block1 = maxpool(block1, kernel_size=(2,2), stride=[1, 2, 2, 1], padding='SAME')
# we use conv-bn-relu in DENSE block (different from paper)
# dropout is taken out of the block
with tf.variable_scope('block2') as scope:
block2 = dense_block_cbr(block1, num=4, num_kernels=16, kernel_size=(3, 3), drop=None)
block2 = maxpool(block2, kernel_size=(2, 2), stride=[1, 2, 2, 1], padding='SAME')
with tf.variable_scope('block3') as scope:
block3 = dense_block_cbr(block2, num=4, num_kernels=24, kernel_size=(3, 3), drop=None)
block3 = dropout(block3, keep=0.9)
block3 = maxpool(block3, kernel_size=(2,2), stride=[1, 2, 2, 1], padding='SAME')
with tf.variable_scope('block4') as scope:
block4 = dense_block_cbr(block3, num=4, num_kernels=32, kernel_size=(3, 3), drop=None)
block4 = conv2d_bn_relu(block4, num_kernels=num_class, kernel_size=(1,1), stride=[1, 1, 1, 1], padding='SAME')
block4 = dropout(block4, keep=0.9)
block4 = avgpool(block4, is_global_pool=True)
logit = block4
return logit
# construct the graph here
make_net = DenseNet_3 #LeNet_0 #LeNet_1
logit = make_net(input_shape =(height, width, channel), output_shape=(num_class))
data = tf.get_default_graph().get_tensor_by_name('input:0')
label = tf.placeholder(dtype=tf.int32, shape=[None])
prob = tf.nn.softmax(logit)
l2 = l2_regulariser(decay=0.0005)
loss = cross_entropy(logit, label)
metric = accuracy(prob, label)
Question 3¶
What does your final architecture look like? (Type of model, layers, sizes, connectivity, etc.) For reference on how to build a deep neural network using TensorFlow, see Deep Neural Network in TensorFlow from the classroom.
Answer: Our network is shown in table below. It as about 27 million MACs (multiply-accumulation operation counts) and has test accuracy of abut 99.68%. The input is 32x32 3-channel input. Below table shows the network structure.
img1=Image(filename="/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/advance/docs/000.png")
img2=Image(filename="/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/advance/docs/001.png")
display(img1, img2)
Our network has the 3 following parts:
[Preprocessing]
- the input is first normalisalised by bn layer. Then it is "preprocessed" by 3x3 and 1x1 conv layers with learnable parameteric ReLU as activations. This is inspired by [1].
[Feature Extraction]
the first block1(conv1) is simply 32 5x5 convolutions in "conv-bn-relu" format. Max pooling is used to reduce the size of the activation.
The next three block block2,3,4 are modified "Dense block", see [2]. Each "Dense block" consists of 4 concatenations of 3x3 convolutions in "conv-bn-relu" format. Note that unlike the paper, dropout is not applied in the block. Instead, we use droupout after the block for regularisation.
[Classification]
- Finally for classification, we apply 1x1 conv and global average pooling.
One uqiue feature of our network is the use of "Dense block" that allows input from lower layers to be "shortcut" to the upper layers. We think this is a way to use multi-scale features. For loss function, we use cross entropy for gradient descent. l2 weight regularisation is also applied.
[1]"Systematic evaluation of CNN advances on the ImageNet"-Dmytro Mishkin, Nikolay Sergievskiy, Jiri Matas, ARXIV 2016 (https://arxiv.org/abs/1606.02228) [2] "Densely Connected Convolutional Networks" - Gao Huang, Zhuang Liu, Kilian Q. Weinberger, Laurens van der Maaten, Arxiv 2016
### Train your model here.
### Feel free to use as many code cells as needed.
# changing of learning rate
def schdule_by_step( r, steps=(0,100), items=(0.1,0.01)):
item = items[0]
N=len(steps)
for n in range(N):
if r >= steps[n]:
item = items[n]
return item
#for testing and validation
def test_net( datas, labels, batch_size, data, label, loss, metric, sess):
num = len(datas)
all_loss = 0
all_acc = 0
all = 0
for n in range(0, num, batch_size):
#print('\r evaluating .... %d/%d' % (n, num), end='', flush=True)
start = n
end = start+batch_size if start+batch_size<=num else num
batch_datas = datas [start:end]
batch_labels = labels [start:end]
fd = {data: batch_datas, label: batch_labels, IS_TRAIN_PHASE : False}
test_loss, test_acc = sess.run([loss, metric], feed_dict=fd)
a = end-start
all += a
all_loss += a*test_loss
all_acc += a*test_acc
assert(all==num)
loss = all_loss/all
acc = all_acc/all
return loss, acc
#solver
epoch_log = 2
max_run = 9
batch_size = 128 #256 #96 384 #128
steps = (0, 3, 6, 8)
rates = (0.1, 0.01, 0.001, 0.0001)
learning_rate = tf.placeholder(tf.float32, shape=[])
solver = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=0.9)
solver_step = solver.minimize(loss+l2)
# start training here ------------------------------------------------
print ('start training')
sess.run(tf.global_variables_initializer(), feed_dict = {IS_TRAIN_PHASE : True } )
saver = tf.train.Saver()
#writer = tf.summary.FileWriter(out_dir + '/tf', graph=tf.get_default_graph())
# keep a log
print('')
print(' run epoch iter rate | train_loss (acc) | valid_loss (acc) | time ')
print('----------------------------------------------------------------------------------------------')
tic = timer()
iter = 0
for r in range(max_run):
rate = schdule_by_step(r, steps=steps, items=rates)
argument_images, argument_labels = shuffle_data_uniform(train_images, train_labels, num_class, num_per_class=num_per_class)
argument_images = make_perturb_images(argument_images, keep=keep)
num_argument = len(argument_images)
N = max(num_argument//batch_size-1,1)
#iter_log = round(float(num_train) / float(num_argument) * float(N))
iter_log = max(round(float( epoch_log *num_train ) / float(batch_size)),1)
for n in range(N):
iter = iter + 1
run = r + float(n)/float(N)
epoch = float(iter*batch_size)/float(num_train)
batch_datas, batch_labels = generate_train_batch_next( argument_images, argument_labels, n, batch_size )
fd = {data: batch_datas, label: batch_labels, learning_rate: rate, IS_TRAIN_PHASE : True }
_, batch_loss, batch_acc, = sess.run([solver_step, loss, metric ],feed_dict=fd)
print('\r%4.1f %5.1f %05d %f | %f (%f) ' %
(run, epoch, iter, rate, batch_loss, batch_acc), end='', flush=True)
#do validation here!
if iter%iter_log==0 or (r==max_run-1 and n==N-1):
toc = timer()
sec_pass = toc - tic
min_pass = sec_pass/60.
#validation
val_loss, val_acc = test_net(valid_images, valid_labels, batch_size, data, label, loss, metric, sess)
#print('\r')
print('\r%4.1f %5.1f %05d %f | %f (%f) | %f (%f) | %4.1f min' %
(run, epoch, iter, rate, batch_loss, batch_acc, val_loss, val_acc, min_pass ), end='\n',flush=True)
pass
# save intermediate checkpoint
# saver.save(sess, out_dir + '/check_points/%06d.ckpt'%r) #iter
#final test! ------------------------------------------
# save final checkpoint
os.makedirs (OUT_DIR + '/check_points/final.ckpt',exist_ok=True)
saver.save(sess, OUT_DIR + '/check_points/final.ckpt')
print('\n')
print('** evaluation on test set **' )
test_loss, test_acc = test_net(test_images, test_labels, batch_size, data, label, loss, metric, sess)
print('test_loss=%f (test_acc=%f)' % ( test_loss, test_acc))
Question 4¶
How did you train your model? (Type of optimizer, batch size, epochs, hyperparameters, etc.)
Answer:
The basic idea is to create "infinite" argumented train data by generating new data during the training epoch. We think that if the train data is always changing, the network will be less prone to overfitting. But we have to make sure that the data change is not too big which will otherwise cause "jumps" in the training loss.
Hence, we divide the training epoches into R runs. For each run, we generate new data by the formula: "train_data_of_current_run = 20% of original_data + 80% of new_augmented_data"
Then, in each run, we have to decide E number of epoch to execute. From our experiments, it turns out that R=9 and E=24 work well. Total number of training epoches is about 9x24 = 216.
To prevent imbalance class problem, we use equal number of samples per class by random sampling during augmentation as mentioned above. We use 2000 samples per class.
For optimizer, we use stochastic gradient descent (sgd) with momentum=0.9. The learning rate is stepped from 0.1, 0.01, 0.001, 0.0001 at rounds 0, 3, 6, 8 respectively. Batch size should be preferably greater than the number of class (43) and we choose to use batch size = 128.
Finally we achieve 99.6% on the test set. We note tha due to different random seed, the results is not exactly the same when we run our code. An example of our loss curve is shown.
img1=Image(filename="/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/advance/docs/002.png")
img2=Image(filename="/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/advance/docs/003.png")
img3=Image(filename="/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/advance/docs/004.png")
display(img1,img2,img3)
Question 5¶
What approach did you take in coming up with a solution to this problem? It may have been a process of trial and error, in which case, outline the steps you took to get to the final solution and why you chose those steps. Perhaps your solution involved an already well known implementation or architecture. In this case, discuss why you think this is suitable for the current problem.
Answer:
Here are our observations:
(1) Establish a target accuracy we want to achive. From searching the internet[3], we know that:
- human performance: 98.84%
- industrial performance: 99.82%
- works by others: 99.10% to 99.67%
Hence we think that a reasonable well trained system should be about 99.5% accuracy.
(2) Train a simple conv net to establish baseline results. Our experiments shows that
- simple LeNet + no dropout + no augmentation: ~90%
- simple LeNet + dropout + no augmentation: ~93%
- simple LeNet + dropout + augmentation: ~95%
- simple LeNet + dropout + augmentation + preprocessing (whitening): ~97%
(Note: simple LeNet = 108_5x5conv_relu-maxpool, 108_5x5conv_relu-maxpool, 100_dense_relu, 100_dense_relu, 43_dense)
From these results, we think that data agumentation and preporcessing could be the key to get higher results. As for newtork structure, LeNet performance seems to be saturated. We have tried Vgg and Denset and decide to use Densenet as our final solution.
(3) Study the data. We confirm that illumination is a key problem.
Based on the observations above, we try the following:
(1) Select a network and tune it.
We choose "Densenet" becuase we think it connect the lower layers to the top. But we did not do much experiments to see how important this is. We try different number of concatentations, different number of kernels and dropput parameters. We find that the orginal "Densenet" is very sensitive to dropout (maybe because the dropout can propagate and get magnified within the block?). Hence we move the droupout out of the block.
(2) Once we fix the network, we choose pre-processing. We try whitening, LRN but find that they are quite slow and doesn't really imporve the results once we have large data argumentation. Hence we decide to got for something simpler, i.e. just include the preprocessing inside the network.
(3) Lastly we try different arugmentations. We find that too much argumentation is no good. we have to limit to the correct rotation, scale, translate, perspective distrortion, brighness, contrast, satuation, etc. We note that final results is sensitive to argumentation.
For each experiment, we carefully record the loss curve. We try different learning rates and weight reqularisation, including L1, L2 and L1+L2. On the whole, it is very much trial and error. Speed of developemnt and training is a key for accuracy improvement. We can't improve results unless we try enough experiments!
[3] Our blogs:
Step 3: Test a Model on New Images¶
Take several pictures of traffic signs that you find on the web or around you (at least five), and run them through your classifier on your computer to produce example results. The classifier might not recognize some local signs but it could prove interesting nonetheless.
You may find signnames.csv useful as it contains mappings from the class id (integer) to the actual sign name.
Implementation¶
Use the code cell (or multiple code cells, if necessary) to implement the first step of your project. Once you have completed your implementation and are satisfied with the results, be sure to thoroughly answer the questions that follow.
### Load the images and plot them here.
### Feel free to use as many code cells as needed.
test_files=['0004.jpg', #normal
'0000.jpg', #normal
'0007.jpg', #occluded with snow
'0006.jpg', #small
'0005.jpg', #not in class
]
test_rois =[(54,180,125,260),(160,430,207,469),(181,32,321,142),(226,65,242,78 ),(388,408,700,676)]
test_label=[13,38,3,40,14]
num=len(test_files)
print('num=%d'%num)
# crop roi to 32x32
results_image = 255. * np.ones(shape=(1 * height, num* width, channel),dtype=np.float32)
results_image1 = 255. * np.ones(shape=(1 * 320, num* 320, channel),dtype=np.float32)
crops = np.zeros(shape=(num,height,width,channel),dtype=np.float32)
for n in range(num):
img = cv2.imread(DATA_DIR+'/extra/' + test_files[n], 1)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB).astype(np.float32)
x1, y1, x2, y2 = test_rois[n]
crop = cv2.resize(img[y1:y2, x1:x2, :], (0, 0), fx=32. / (x2 - x1), fy=32. / (y2 - y1),
interpolation=cv2.INTER_CUBIC)
crop = np.clip(crop,0,255)
crops[n]=crop
insert_subimage(results_image, crop, 0, n*width)
#mak roi and show
H,W,C=img.shape
S=max(H,W)
f=320./S
norm_img = cv2.resize(img, (0, 0), fx=f, fy=f, interpolation=cv2.INTER_CUBIC)
cv2.rectangle(norm_img, (round(f*x1), round(f*y1)), (round(f*x2), round(f*y2)), (255,255,0), 3)
insert_subimage(results_image1, norm_img, 0, n*320)
#cv2.imshow('crop', crop)
#cv2.imshow('img', img)
#cv2.waitKey(0)
cv2.imwrite(OUT_DIR+'/extra_crops.jpg', cv2.cvtColor(results_image, cv2.COLOR_BGR2RGB) )
cv2.imwrite(OUT_DIR+'/extra_marked.jpg', cv2.cvtColor(results_image1, cv2.COLOR_BGR2RGB) )
plt.rcParams["figure.figsize"] = (25,25)
plt.imshow(results_image.astype(np.uint8))
plt.axis('off')
plt.show()
plt.imshow(results_image1.astype(np.uint8))
plt.axis('off')
plt.show()
Question 6¶
Choose five candidate images of traffic signs and provide them in the report. Are there any particular qualities of the image(s) that might make classification difficult? It could be helpful to plot the images in the notebook.
Answer: We downloaded 5 images from the internet and determine the roi region of the traffic sign by hand. The images are:
img1=Image(filename="/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/advance/docs/006.png")
display(img1)
### Run the predictions here.
### Feel free to use as many code cells as needed.
#remember to preprocess!
new_images = crops #preprocess(crops)
# load trained classifier
saver = tf.train.Saver()
saver.restore(sess, OUT_DIR + '/check_points/final.ckpt')
print('** test on extra **')
fd = {data: new_images, IS_TRAIN_PHASE: False}
test_prob = sess.run(prob, feed_dict=fd)
print('see printout of results in the next jupyter cell!')
print('sucess')
Question 7¶
Is your model able to perform equally well on captured pictures when compared to testing on the dataset? The simplest way to do this check the accuracy of the predictions. For example, if the model predicted 1 out of 5 signs correctly, it's 20% accurate.
NOTE: You could check the accuracy manually by using signnames.csv (same directory). This file has a mapping from the class id (0-42) to the corresponding sign name. So, you could take the class id the model outputs, lookup the name in signnames.csv and see if it matches the sign from the image.
Answer:
Accuracy is 80%, 4 out of 5 is correct. The results are:
img1=Image(filename="/root/share/docs/git/hengck23-udacity/udacity-driverless-car-nd-p2/advance/docs/007.png")
display(img1)
### Visualize the softmax probabilities here.
### Feel free to use as many code cells as needed.
#show results
f=10
results_image = 255. * np.ones(shape=(5*(f*height + f*8), 6*f*width, channel), dtype=np.float32)
for n in range(num):
crop = crops[n]
c_hat=test_label[n]
c_hat_label = classnames[c_hat] if c_hat>=0 else 'NIL'
print('n=%d: true = %02d:%s' % (n, c_hat,c_hat_label))
#crop = cv2.resize(crop, (0, 0), fx=f, fy=f, interpolation=cv2.INTER_NN)
crop = crop.repeat(f, axis=0).repeat(f, axis=1)
insert_subimage(results_image, crop, n * (f*height + f*8), 0)
cv2.putText(results_image, '%02d:%s%s' % (c_hat,c_hat_label[0:15], '...' if len(classnames[c_hat])>15 else ''),
(5, (n+1) * (f*height + f*8)-50), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 0), 2)
p = test_prob[n]
idx = np.argsort(p)[::-1]
for k in range(5):
c = int(idx[k])
label_image = get_label_image(c)
#label_image = cv2.resize(label_image, (0, 0), fx=f, fy=f, interpolation=cv2.INTER_NN)
label_image = label_image.repeat(f, axis=0).repeat(f, axis=1)
insert_subimage(results_image, label_image, n * (f*height + f*8), (k + 1) * f*width)
print('\ttop%d: %f %02d:%s' % (k, p[c], c, classnames[c]))
cv2.putText(results_image, '%02d:%s%s' % (c, classnames[c][0:15], '...' if len(classnames[c])>15 else ''),
(5+(k + 1) * f*width, (n+1) * (f*height + f*8)-50), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 0), 2)
cv2.putText(results_image, 'top%d: %f' % (k, p[c]),
(5+(k + 1) * f*width, (n+1) * (f*height + f*8)-20), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 0), 2)
print('')
print('')
print('')
print('**visual results**: input, followed by top-5')
cv2.imwrite(OUT_DIR+'/extra_predictions.jpg', cv2.cvtColor(results_image, cv2.COLOR_BGR2RGB) )
plt.rcParams["figure.figsize"] = (30,30)
plt.imshow(results_image.astype(np.uint8))
plt.axis('off')
plt.show()
Question 8¶
Use the model's softmax probabilities to visualize the certainty of its predictions, tf.nn.top_k could prove helpful here. Which predictions is the model certain of? Uncertain? If the model was incorrect in its initial prediction, does the correct prediction appear in the top k? (k should be 5 at most)
tf.nn.top_k will return the values and indices (class ids) of the top k predictions. So if k=3, for each sign, it'll return the 3 largest probabilities (out of a possible 43) and the correspoding class ids.
Take this numpy array as an example:
# (5, 6) array
a = np.array([[ 0.24879643, 0.07032244, 0.12641572, 0.34763842, 0.07893497,
0.12789202],
[ 0.28086119, 0.27569815, 0.08594638, 0.0178669 , 0.18063401,
0.15899337],
[ 0.26076848, 0.23664738, 0.08020603, 0.07001922, 0.1134371 ,
0.23892179],
[ 0.11943333, 0.29198961, 0.02605103, 0.26234032, 0.1351348 ,
0.16505091],
[ 0.09561176, 0.34396535, 0.0643941 , 0.16240774, 0.24206137,
0.09155967]])Running it through sess.run(tf.nn.top_k(tf.constant(a), k=3)) produces:
TopKV2(values=array([[ 0.34763842, 0.24879643, 0.12789202],
[ 0.28086119, 0.27569815, 0.18063401],
[ 0.26076848, 0.23892179, 0.23664738],
[ 0.29198961, 0.26234032, 0.16505091],
[ 0.34396535, 0.24206137, 0.16240774]]), indices=array([[3, 0, 5],
[0, 1, 4],
[0, 5, 1],
[1, 3, 5],
[1, 4, 3]], dtype=int32))Looking just at the first row we get [ 0.34763842, 0.24879643, 0.12789202], you can confirm these are the 3 largest probabilities in a. You'll also notice [3, 0, 5] are the corresponding indices.
Answer:
Our classifier is correct for "normal" and "small" cases. What it is correct the top-1 score is very high.
Our classifier is wrong for "occluded with snow" case. Here, the top score is only 0.35. The correct label is not within the top 5 at all. This is indeed a difficult case.
For the last case, the test image is not in the training sample. But the classifier makes a reasonable prediction. The "stop sign" is indeed the closest match to the test image (the color and hexagon shape). The top-1 score may seem to be too high however.
Note: Once you have completed all of the code implementations and successfully answered each question above, you may finalize your work by exporting the iPython Notebook as an HTML document. You can do this by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.